1c942fddfSThomas Gleixner // SPDX-License-Identifier: GPL-2.0-or-later 2133ff0eaSJérôme Glisse /* 3133ff0eaSJérôme Glisse * Copyright 2013 Red Hat Inc. 4133ff0eaSJérôme Glisse * 5f813f219SJérôme Glisse * Authors: Jérôme Glisse <jglisse@redhat.com> 6133ff0eaSJérôme Glisse */ 7133ff0eaSJérôme Glisse /* 8133ff0eaSJérôme Glisse * Refer to include/linux/hmm.h for information about heterogeneous memory 9133ff0eaSJérôme Glisse * management or HMM for short. 10133ff0eaSJérôme Glisse */ 11133ff0eaSJérôme Glisse #include <linux/mm.h> 12133ff0eaSJérôme Glisse #include <linux/hmm.h> 13858b54daSJérôme Glisse #include <linux/init.h> 14da4c3c73SJérôme Glisse #include <linux/rmap.h> 15da4c3c73SJérôme Glisse #include <linux/swap.h> 16133ff0eaSJérôme Glisse #include <linux/slab.h> 17133ff0eaSJérôme Glisse #include <linux/sched.h> 184ef589dcSJérôme Glisse #include <linux/mmzone.h> 194ef589dcSJérôme Glisse #include <linux/pagemap.h> 20da4c3c73SJérôme Glisse #include <linux/swapops.h> 21da4c3c73SJérôme Glisse #include <linux/hugetlb.h> 224ef589dcSJérôme Glisse #include <linux/memremap.h> 237b2d55d2SJérôme Glisse #include <linux/jump_label.h> 2455c0ece8SJérôme Glisse #include <linux/dma-mapping.h> 25c0b12405SJérôme Glisse #include <linux/mmu_notifier.h> 264ef589dcSJérôme Glisse #include <linux/memory_hotplug.h> 274ef589dcSJérôme Glisse 284ef589dcSJérôme Glisse #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT) 29133ff0eaSJérôme Glisse 306b368cd4SJérôme Glisse #if IS_ENABLED(CONFIG_HMM_MIRROR) 31c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops; 32c0b12405SJérôme Glisse 33704f3f2cSJérôme Glisse static inline struct hmm *mm_get_hmm(struct mm_struct *mm) 34133ff0eaSJérôme Glisse { 35c0b12405SJérôme Glisse struct hmm *hmm = READ_ONCE(mm->hmm); 36704f3f2cSJérôme Glisse 37704f3f2cSJérôme Glisse if (hmm && kref_get_unless_zero(&hmm->kref)) 38704f3f2cSJérôme Glisse return hmm; 39704f3f2cSJérôme Glisse 40704f3f2cSJérôme Glisse return NULL; 41704f3f2cSJérôme Glisse } 42704f3f2cSJérôme Glisse 43704f3f2cSJérôme Glisse /** 44704f3f2cSJérôme Glisse * hmm_get_or_create - register HMM against an mm (HMM internal) 45704f3f2cSJérôme Glisse * 46704f3f2cSJérôme Glisse * @mm: mm struct to attach to 47704f3f2cSJérôme Glisse * Returns: returns an HMM object, either by referencing the existing 48704f3f2cSJérôme Glisse * (per-process) object, or by creating a new one. 49704f3f2cSJérôme Glisse * 50704f3f2cSJérôme Glisse * This is not intended to be used directly by device drivers. If mm already 51704f3f2cSJérôme Glisse * has an HMM struct then it get a reference on it and returns it. Otherwise 52704f3f2cSJérôme Glisse * it allocates an HMM struct, initializes it, associate it with the mm and 53704f3f2cSJérôme Glisse * returns it. 54704f3f2cSJérôme Glisse */ 55704f3f2cSJérôme Glisse static struct hmm *hmm_get_or_create(struct mm_struct *mm) 56704f3f2cSJérôme Glisse { 57704f3f2cSJérôme Glisse struct hmm *hmm = mm_get_hmm(mm); 58c0b12405SJérôme Glisse bool cleanup = false; 59133ff0eaSJérôme Glisse 60c0b12405SJérôme Glisse if (hmm) 61c0b12405SJérôme Glisse return hmm; 62c0b12405SJérôme Glisse 63c0b12405SJérôme Glisse hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); 64c0b12405SJérôme Glisse if (!hmm) 65c0b12405SJérôme Glisse return NULL; 66a3e0d41cSJérôme Glisse init_waitqueue_head(&hmm->wq); 67c0b12405SJérôme Glisse INIT_LIST_HEAD(&hmm->mirrors); 68c0b12405SJérôme Glisse init_rwsem(&hmm->mirrors_sem); 69c0b12405SJérôme Glisse hmm->mmu_notifier.ops = NULL; 70da4c3c73SJérôme Glisse INIT_LIST_HEAD(&hmm->ranges); 71a3e0d41cSJérôme Glisse mutex_init(&hmm->lock); 72704f3f2cSJérôme Glisse kref_init(&hmm->kref); 73a3e0d41cSJérôme Glisse hmm->notifiers = 0; 74a3e0d41cSJérôme Glisse hmm->dead = false; 75c0b12405SJérôme Glisse hmm->mm = mm; 76c0b12405SJérôme Glisse 77c0b12405SJérôme Glisse spin_lock(&mm->page_table_lock); 78c0b12405SJérôme Glisse if (!mm->hmm) 79c0b12405SJérôme Glisse mm->hmm = hmm; 80c0b12405SJérôme Glisse else 81c0b12405SJérôme Glisse cleanup = true; 82c0b12405SJérôme Glisse spin_unlock(&mm->page_table_lock); 83c0b12405SJérôme Glisse 8486a2d598SRalph Campbell if (cleanup) 8586a2d598SRalph Campbell goto error; 8686a2d598SRalph Campbell 8786a2d598SRalph Campbell /* 8886a2d598SRalph Campbell * We should only get here if hold the mmap_sem in write mode ie on 8986a2d598SRalph Campbell * registration of first mirror through hmm_mirror_register() 9086a2d598SRalph Campbell */ 9186a2d598SRalph Campbell hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; 9286a2d598SRalph Campbell if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) 9386a2d598SRalph Campbell goto error_mm; 94c0b12405SJérôme Glisse 95704f3f2cSJérôme Glisse return hmm; 9686a2d598SRalph Campbell 9786a2d598SRalph Campbell error_mm: 9886a2d598SRalph Campbell spin_lock(&mm->page_table_lock); 9986a2d598SRalph Campbell if (mm->hmm == hmm) 10086a2d598SRalph Campbell mm->hmm = NULL; 10186a2d598SRalph Campbell spin_unlock(&mm->page_table_lock); 10286a2d598SRalph Campbell error: 10386a2d598SRalph Campbell kfree(hmm); 10486a2d598SRalph Campbell return NULL; 105133ff0eaSJérôme Glisse } 106133ff0eaSJérôme Glisse 1076d7c3cdeSJason Gunthorpe static void hmm_free_rcu(struct rcu_head *rcu) 1086d7c3cdeSJason Gunthorpe { 1096d7c3cdeSJason Gunthorpe kfree(container_of(rcu, struct hmm, rcu)); 1106d7c3cdeSJason Gunthorpe } 1116d7c3cdeSJason Gunthorpe 112704f3f2cSJérôme Glisse static void hmm_free(struct kref *kref) 113704f3f2cSJérôme Glisse { 114704f3f2cSJérôme Glisse struct hmm *hmm = container_of(kref, struct hmm, kref); 115704f3f2cSJérôme Glisse struct mm_struct *mm = hmm->mm; 116704f3f2cSJérôme Glisse 117704f3f2cSJérôme Glisse mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm); 118704f3f2cSJérôme Glisse 119704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 120704f3f2cSJérôme Glisse if (mm->hmm == hmm) 121704f3f2cSJérôme Glisse mm->hmm = NULL; 122704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 123704f3f2cSJérôme Glisse 1246d7c3cdeSJason Gunthorpe mmu_notifier_call_srcu(&hmm->rcu, hmm_free_rcu); 125704f3f2cSJérôme Glisse } 126704f3f2cSJérôme Glisse 127704f3f2cSJérôme Glisse static inline void hmm_put(struct hmm *hmm) 128704f3f2cSJérôme Glisse { 129704f3f2cSJérôme Glisse kref_put(&hmm->kref, hmm_free); 130704f3f2cSJérôme Glisse } 131704f3f2cSJérôme Glisse 132133ff0eaSJérôme Glisse void hmm_mm_destroy(struct mm_struct *mm) 133133ff0eaSJérôme Glisse { 134704f3f2cSJérôme Glisse struct hmm *hmm; 135704f3f2cSJérôme Glisse 136704f3f2cSJérôme Glisse spin_lock(&mm->page_table_lock); 137704f3f2cSJérôme Glisse hmm = mm_get_hmm(mm); 138704f3f2cSJérôme Glisse mm->hmm = NULL; 139704f3f2cSJérôme Glisse if (hmm) { 140704f3f2cSJérôme Glisse hmm->mm = NULL; 141a3e0d41cSJérôme Glisse hmm->dead = true; 142704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 143704f3f2cSJérôme Glisse hmm_put(hmm); 144704f3f2cSJérôme Glisse return; 145704f3f2cSJérôme Glisse } 146704f3f2cSJérôme Glisse 147704f3f2cSJérôme Glisse spin_unlock(&mm->page_table_lock); 148133ff0eaSJérôme Glisse } 149c0b12405SJérôme Glisse 150a3e0d41cSJérôme Glisse static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) 151c0b12405SJérôme Glisse { 1526d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 153c0b12405SJérôme Glisse struct hmm_mirror *mirror; 154da4c3c73SJérôme Glisse struct hmm_range *range; 155da4c3c73SJérôme Glisse 1566d7c3cdeSJason Gunthorpe /* Bail out if hmm is in the process of being freed */ 1576d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 1586d7c3cdeSJason Gunthorpe return; 1596d7c3cdeSJason Gunthorpe 160a3e0d41cSJérôme Glisse /* Report this HMM as dying. */ 161a3e0d41cSJérôme Glisse hmm->dead = true; 162da4c3c73SJérôme Glisse 163a3e0d41cSJérôme Glisse /* Wake-up everyone waiting on any range. */ 164a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 165085ea250SRalph Campbell list_for_each_entry(range, &hmm->ranges, list) 166da4c3c73SJérôme Glisse range->valid = false; 167a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 168a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 169e1401513SRalph Campbell 170e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 171e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror, 172e1401513SRalph Campbell list); 173e1401513SRalph Campbell while (mirror) { 174e1401513SRalph Campbell list_del_init(&mirror->list); 175e1401513SRalph Campbell if (mirror->ops->release) { 176e1401513SRalph Campbell /* 177085ea250SRalph Campbell * Drop mirrors_sem so the release callback can wait 178085ea250SRalph Campbell * on any pending work that might itself trigger a 179085ea250SRalph Campbell * mmu_notifier callback and thus would deadlock with 180085ea250SRalph Campbell * us. 181e1401513SRalph Campbell */ 182e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 183e1401513SRalph Campbell mirror->ops->release(mirror); 184e1401513SRalph Campbell down_write(&hmm->mirrors_sem); 185e1401513SRalph Campbell } 186e1401513SRalph Campbell mirror = list_first_entry_or_null(&hmm->mirrors, 187e1401513SRalph Campbell struct hmm_mirror, list); 188e1401513SRalph Campbell } 189e1401513SRalph Campbell up_write(&hmm->mirrors_sem); 190704f3f2cSJérôme Glisse 191704f3f2cSJérôme Glisse hmm_put(hmm); 192e1401513SRalph Campbell } 193e1401513SRalph Campbell 19493065ac7SMichal Hocko static int hmm_invalidate_range_start(struct mmu_notifier *mn, 195a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 196c0b12405SJérôme Glisse { 1976d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 198a3e0d41cSJérôme Glisse struct hmm_mirror *mirror; 199ec131b2dSJérôme Glisse struct hmm_update update; 200a3e0d41cSJérôme Glisse struct hmm_range *range; 201a3e0d41cSJérôme Glisse int ret = 0; 202c0b12405SJérôme Glisse 2036d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 2046d7c3cdeSJason Gunthorpe return 0; 205c0b12405SJérôme Glisse 206a3e0d41cSJérôme Glisse update.start = nrange->start; 207a3e0d41cSJérôme Glisse update.end = nrange->end; 208ec131b2dSJérôme Glisse update.event = HMM_UPDATE_INVALIDATE; 209dfcd6660SJérôme Glisse update.blockable = mmu_notifier_range_blockable(nrange); 210a3e0d41cSJérôme Glisse 211dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 212a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 213a3e0d41cSJérôme Glisse else if (!mutex_trylock(&hmm->lock)) { 214a3e0d41cSJérôme Glisse ret = -EAGAIN; 215a3e0d41cSJérôme Glisse goto out; 216a3e0d41cSJérôme Glisse } 217a3e0d41cSJérôme Glisse hmm->notifiers++; 218a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 219a3e0d41cSJérôme Glisse if (update.end < range->start || update.start >= range->end) 220a3e0d41cSJérôme Glisse continue; 221a3e0d41cSJérôme Glisse 222a3e0d41cSJérôme Glisse range->valid = false; 223a3e0d41cSJérôme Glisse } 224a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 225a3e0d41cSJérôme Glisse 226dfcd6660SJérôme Glisse if (mmu_notifier_range_blockable(nrange)) 227a3e0d41cSJérôme Glisse down_read(&hmm->mirrors_sem); 228a3e0d41cSJérôme Glisse else if (!down_read_trylock(&hmm->mirrors_sem)) { 229a3e0d41cSJérôme Glisse ret = -EAGAIN; 230a3e0d41cSJérôme Glisse goto out; 231a3e0d41cSJérôme Glisse } 232a3e0d41cSJérôme Glisse list_for_each_entry(mirror, &hmm->mirrors, list) { 233a3e0d41cSJérôme Glisse int ret; 234a3e0d41cSJérôme Glisse 235a3e0d41cSJérôme Glisse ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update); 236085ea250SRalph Campbell if (!update.blockable && ret == -EAGAIN) 237085ea250SRalph Campbell break; 238a3e0d41cSJérôme Glisse } 239a3e0d41cSJérôme Glisse up_read(&hmm->mirrors_sem); 240a3e0d41cSJérôme Glisse 241a3e0d41cSJérôme Glisse out: 242704f3f2cSJérôme Glisse hmm_put(hmm); 243704f3f2cSJérôme Glisse return ret; 244c0b12405SJérôme Glisse } 245c0b12405SJérôme Glisse 246c0b12405SJérôme Glisse static void hmm_invalidate_range_end(struct mmu_notifier *mn, 247a3e0d41cSJérôme Glisse const struct mmu_notifier_range *nrange) 248c0b12405SJérôme Glisse { 2496d7c3cdeSJason Gunthorpe struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier); 250c0b12405SJérôme Glisse 2516d7c3cdeSJason Gunthorpe if (!kref_get_unless_zero(&hmm->kref)) 2526d7c3cdeSJason Gunthorpe return; 253c0b12405SJérôme Glisse 254a3e0d41cSJérôme Glisse mutex_lock(&hmm->lock); 255a3e0d41cSJérôme Glisse hmm->notifiers--; 256a3e0d41cSJérôme Glisse if (!hmm->notifiers) { 257a3e0d41cSJérôme Glisse struct hmm_range *range; 258a3e0d41cSJérôme Glisse 259a3e0d41cSJérôme Glisse list_for_each_entry(range, &hmm->ranges, list) { 260a3e0d41cSJérôme Glisse if (range->valid) 261a3e0d41cSJérôme Glisse continue; 262a3e0d41cSJérôme Glisse range->valid = true; 263a3e0d41cSJérôme Glisse } 264a3e0d41cSJérôme Glisse wake_up_all(&hmm->wq); 265a3e0d41cSJérôme Glisse } 266a3e0d41cSJérôme Glisse mutex_unlock(&hmm->lock); 267a3e0d41cSJérôme Glisse 268704f3f2cSJérôme Glisse hmm_put(hmm); 269c0b12405SJérôme Glisse } 270c0b12405SJérôme Glisse 271c0b12405SJérôme Glisse static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { 272e1401513SRalph Campbell .release = hmm_release, 273c0b12405SJérôme Glisse .invalidate_range_start = hmm_invalidate_range_start, 274c0b12405SJérôme Glisse .invalidate_range_end = hmm_invalidate_range_end, 275c0b12405SJérôme Glisse }; 276c0b12405SJérôme Glisse 277c0b12405SJérôme Glisse /* 278c0b12405SJérôme Glisse * hmm_mirror_register() - register a mirror against an mm 279c0b12405SJérôme Glisse * 280c0b12405SJérôme Glisse * @mirror: new mirror struct to register 281c0b12405SJérôme Glisse * @mm: mm to register against 282085ea250SRalph Campbell * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments 283c0b12405SJérôme Glisse * 284c0b12405SJérôme Glisse * To start mirroring a process address space, the device driver must register 285c0b12405SJérôme Glisse * an HMM mirror struct. 286c0b12405SJérôme Glisse * 287c0b12405SJérôme Glisse * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! 288c0b12405SJérôme Glisse */ 289c0b12405SJérôme Glisse int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) 290c0b12405SJérôme Glisse { 291c0b12405SJérôme Glisse /* Sanity check */ 292c0b12405SJérôme Glisse if (!mm || !mirror || !mirror->ops) 293c0b12405SJérôme Glisse return -EINVAL; 294c0b12405SJérôme Glisse 295704f3f2cSJérôme Glisse mirror->hmm = hmm_get_or_create(mm); 296c0b12405SJérôme Glisse if (!mirror->hmm) 297c0b12405SJérôme Glisse return -ENOMEM; 298c0b12405SJérôme Glisse 299c0b12405SJérôme Glisse down_write(&mirror->hmm->mirrors_sem); 300c0b12405SJérôme Glisse list_add(&mirror->list, &mirror->hmm->mirrors); 301c0b12405SJérôme Glisse up_write(&mirror->hmm->mirrors_sem); 302c0b12405SJérôme Glisse 303c0b12405SJérôme Glisse return 0; 304c0b12405SJérôme Glisse } 305c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_register); 306c0b12405SJérôme Glisse 307c0b12405SJérôme Glisse /* 308c0b12405SJérôme Glisse * hmm_mirror_unregister() - unregister a mirror 309c0b12405SJérôme Glisse * 310085ea250SRalph Campbell * @mirror: mirror struct to unregister 311c0b12405SJérôme Glisse * 312c0b12405SJérôme Glisse * Stop mirroring a process address space, and cleanup. 313c0b12405SJérôme Glisse */ 314c0b12405SJérôme Glisse void hmm_mirror_unregister(struct hmm_mirror *mirror) 315c0b12405SJérôme Glisse { 316704f3f2cSJérôme Glisse struct hmm *hmm = READ_ONCE(mirror->hmm); 317c0b12405SJérôme Glisse 318704f3f2cSJérôme Glisse if (hmm == NULL) 319c01cbba2SJérôme Glisse return; 320c01cbba2SJérôme Glisse 321c0b12405SJérôme Glisse down_write(&hmm->mirrors_sem); 322e1401513SRalph Campbell list_del_init(&mirror->list); 323704f3f2cSJérôme Glisse /* To protect us against double unregister ... */ 324c01cbba2SJérôme Glisse mirror->hmm = NULL; 325c0b12405SJérôme Glisse up_write(&hmm->mirrors_sem); 326c01cbba2SJérôme Glisse 327704f3f2cSJérôme Glisse hmm_put(hmm); 328c0b12405SJérôme Glisse } 329c0b12405SJérôme Glisse EXPORT_SYMBOL(hmm_mirror_unregister); 330da4c3c73SJérôme Glisse 33174eee180SJérôme Glisse struct hmm_vma_walk { 33274eee180SJérôme Glisse struct hmm_range *range; 333992de9a8SJérôme Glisse struct dev_pagemap *pgmap; 33474eee180SJérôme Glisse unsigned long last; 33574eee180SJérôme Glisse bool fault; 33674eee180SJérôme Glisse bool block; 33774eee180SJérôme Glisse }; 33874eee180SJérôme Glisse 3392aee09d8SJérôme Glisse static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, 3402aee09d8SJérôme Glisse bool write_fault, uint64_t *pfn) 34174eee180SJérôme Glisse { 3429b1ae605SKuehling, Felix unsigned int flags = FAULT_FLAG_REMOTE; 34374eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 344f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 34574eee180SJérôme Glisse struct vm_area_struct *vma = walk->vma; 34650a7ca3cSSouptick Joarder vm_fault_t ret; 34774eee180SJérôme Glisse 34874eee180SJérôme Glisse flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; 3492aee09d8SJérôme Glisse flags |= write_fault ? FAULT_FLAG_WRITE : 0; 35050a7ca3cSSouptick Joarder ret = handle_mm_fault(vma, addr, flags); 35150a7ca3cSSouptick Joarder if (ret & VM_FAULT_RETRY) 35273231612SJérôme Glisse return -EAGAIN; 35350a7ca3cSSouptick Joarder if (ret & VM_FAULT_ERROR) { 354f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 35574eee180SJérôme Glisse return -EFAULT; 35674eee180SJérôme Glisse } 35774eee180SJérôme Glisse 35873231612SJérôme Glisse return -EBUSY; 35974eee180SJérôme Glisse } 36074eee180SJérôme Glisse 361da4c3c73SJérôme Glisse static int hmm_pfns_bad(unsigned long addr, 362da4c3c73SJérôme Glisse unsigned long end, 363da4c3c73SJérôme Glisse struct mm_walk *walk) 364da4c3c73SJérôme Glisse { 365c719547fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 366c719547fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 367ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 368da4c3c73SJérôme Glisse unsigned long i; 369da4c3c73SJérôme Glisse 370da4c3c73SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 371da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, i++) 372f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_ERROR]; 373da4c3c73SJérôme Glisse 374da4c3c73SJérôme Glisse return 0; 375da4c3c73SJérôme Glisse } 376da4c3c73SJérôme Glisse 3775504ed29SJérôme Glisse /* 3785504ed29SJérôme Glisse * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) 3795504ed29SJérôme Glisse * @start: range virtual start address (inclusive) 3805504ed29SJérôme Glisse * @end: range virtual end address (exclusive) 3812aee09d8SJérôme Glisse * @fault: should we fault or not ? 3822aee09d8SJérôme Glisse * @write_fault: write fault ? 3835504ed29SJérôme Glisse * @walk: mm_walk structure 384085ea250SRalph Campbell * Return: 0 on success, -EBUSY after page fault, or page fault error 3855504ed29SJérôme Glisse * 3865504ed29SJérôme Glisse * This function will be called whenever pmd_none() or pte_none() returns true, 3875504ed29SJérôme Glisse * or whenever there is no page directory covering the virtual address range. 3885504ed29SJérôme Glisse */ 3892aee09d8SJérôme Glisse static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, 3902aee09d8SJérôme Glisse bool fault, bool write_fault, 391da4c3c73SJérôme Glisse struct mm_walk *walk) 392da4c3c73SJérôme Glisse { 39374eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 39474eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 395ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 39663d5066fSJérôme Glisse unsigned long i, page_size; 397da4c3c73SJérôme Glisse 39874eee180SJérôme Glisse hmm_vma_walk->last = addr; 39963d5066fSJérôme Glisse page_size = hmm_range_page_size(range); 40063d5066fSJérôme Glisse i = (addr - range->start) >> range->page_shift; 40163d5066fSJérôme Glisse 40263d5066fSJérôme Glisse for (; addr < end; addr += page_size, i++) { 403f88a1e90SJérôme Glisse pfns[i] = range->values[HMM_PFN_NONE]; 4042aee09d8SJérôme Glisse if (fault || write_fault) { 40574eee180SJérôme Glisse int ret; 406da4c3c73SJérôme Glisse 4072aee09d8SJérôme Glisse ret = hmm_vma_do_fault(walk, addr, write_fault, 4082aee09d8SJérôme Glisse &pfns[i]); 40973231612SJérôme Glisse if (ret != -EBUSY) 41074eee180SJérôme Glisse return ret; 41174eee180SJérôme Glisse } 41274eee180SJérôme Glisse } 41374eee180SJérôme Glisse 41473231612SJérôme Glisse return (fault || write_fault) ? -EBUSY : 0; 4152aee09d8SJérôme Glisse } 4162aee09d8SJérôme Glisse 4172aee09d8SJérôme Glisse static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4182aee09d8SJérôme Glisse uint64_t pfns, uint64_t cpu_flags, 4192aee09d8SJérôme Glisse bool *fault, bool *write_fault) 4202aee09d8SJérôme Glisse { 421f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 422f88a1e90SJérôme Glisse 4232aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) 4242aee09d8SJérôme Glisse return; 4252aee09d8SJérôme Glisse 426023a019aSJérôme Glisse /* 427023a019aSJérôme Glisse * So we not only consider the individual per page request we also 428023a019aSJérôme Glisse * consider the default flags requested for the range. The API can 429023a019aSJérôme Glisse * be use in 2 fashions. The first one where the HMM user coalesce 430023a019aSJérôme Glisse * multiple page fault into one request and set flags per pfns for 431023a019aSJérôme Glisse * of those faults. The second one where the HMM user want to pre- 432023a019aSJérôme Glisse * fault a range with specific flags. For the latter one it is a 433023a019aSJérôme Glisse * waste to have the user pre-fill the pfn arrays with a default 434023a019aSJérôme Glisse * flags value. 435023a019aSJérôme Glisse */ 436023a019aSJérôme Glisse pfns = (pfns & range->pfn_flags_mask) | range->default_flags; 437023a019aSJérôme Glisse 4382aee09d8SJérôme Glisse /* We aren't ask to do anything ... */ 439f88a1e90SJérôme Glisse if (!(pfns & range->flags[HMM_PFN_VALID])) 4402aee09d8SJérôme Glisse return; 441f88a1e90SJérôme Glisse /* If this is device memory than only fault if explicitly requested */ 442f88a1e90SJérôme Glisse if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { 443f88a1e90SJérôme Glisse /* Do we fault on device memory ? */ 444f88a1e90SJérôme Glisse if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { 445f88a1e90SJérôme Glisse *write_fault = pfns & range->flags[HMM_PFN_WRITE]; 446f88a1e90SJérôme Glisse *fault = true; 447f88a1e90SJérôme Glisse } 4482aee09d8SJérôme Glisse return; 4492aee09d8SJérôme Glisse } 450f88a1e90SJérôme Glisse 451f88a1e90SJérôme Glisse /* If CPU page table is not valid then we need to fault */ 452f88a1e90SJérôme Glisse *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); 453f88a1e90SJérôme Glisse /* Need to write fault ? */ 454f88a1e90SJérôme Glisse if ((pfns & range->flags[HMM_PFN_WRITE]) && 455f88a1e90SJérôme Glisse !(cpu_flags & range->flags[HMM_PFN_WRITE])) { 456f88a1e90SJérôme Glisse *write_fault = true; 4572aee09d8SJérôme Glisse *fault = true; 4582aee09d8SJérôme Glisse } 4592aee09d8SJérôme Glisse } 4602aee09d8SJérôme Glisse 4612aee09d8SJérôme Glisse static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, 4622aee09d8SJérôme Glisse const uint64_t *pfns, unsigned long npages, 4632aee09d8SJérôme Glisse uint64_t cpu_flags, bool *fault, 4642aee09d8SJérôme Glisse bool *write_fault) 4652aee09d8SJérôme Glisse { 4662aee09d8SJérôme Glisse unsigned long i; 4672aee09d8SJérôme Glisse 4682aee09d8SJérôme Glisse if (!hmm_vma_walk->fault) { 4692aee09d8SJérôme Glisse *fault = *write_fault = false; 4702aee09d8SJérôme Glisse return; 4712aee09d8SJérôme Glisse } 4722aee09d8SJérôme Glisse 473a3e0d41cSJérôme Glisse *fault = *write_fault = false; 4742aee09d8SJérôme Glisse for (i = 0; i < npages; ++i) { 4752aee09d8SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, 4762aee09d8SJérôme Glisse fault, write_fault); 477a3e0d41cSJérôme Glisse if ((*write_fault)) 4782aee09d8SJérôme Glisse return; 4792aee09d8SJérôme Glisse } 4802aee09d8SJérôme Glisse } 4812aee09d8SJérôme Glisse 4822aee09d8SJérôme Glisse static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, 4832aee09d8SJérôme Glisse struct mm_walk *walk) 4842aee09d8SJérôme Glisse { 4852aee09d8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 4862aee09d8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 4872aee09d8SJérôme Glisse bool fault, write_fault; 4882aee09d8SJérôme Glisse unsigned long i, npages; 4892aee09d8SJérôme Glisse uint64_t *pfns; 4902aee09d8SJérôme Glisse 4912aee09d8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 4922aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 4932aee09d8SJérôme Glisse pfns = &range->pfns[i]; 4942aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 4952aee09d8SJérôme Glisse 0, &fault, &write_fault); 4962aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 4972aee09d8SJérôme Glisse } 4982aee09d8SJérôme Glisse 499f88a1e90SJérôme Glisse static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) 5002aee09d8SJérôme Glisse { 5012aee09d8SJérôme Glisse if (pmd_protnone(pmd)) 5022aee09d8SJérôme Glisse return 0; 503f88a1e90SJérôme Glisse return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | 504f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 505f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 506da4c3c73SJérôme Glisse } 507da4c3c73SJérôme Glisse 508992de9a8SJérôme Glisse static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud) 509992de9a8SJérôme Glisse { 510992de9a8SJérôme Glisse if (!pud_present(pud)) 511992de9a8SJérôme Glisse return 0; 512992de9a8SJérôme Glisse return pud_write(pud) ? range->flags[HMM_PFN_VALID] | 513992de9a8SJérôme Glisse range->flags[HMM_PFN_WRITE] : 514992de9a8SJérôme Glisse range->flags[HMM_PFN_VALID]; 515992de9a8SJérôme Glisse } 516992de9a8SJérôme Glisse 51753f5c3f4SJérôme Glisse static int hmm_vma_handle_pmd(struct mm_walk *walk, 51853f5c3f4SJérôme Glisse unsigned long addr, 51953f5c3f4SJérôme Glisse unsigned long end, 52053f5c3f4SJérôme Glisse uint64_t *pfns, 52153f5c3f4SJérôme Glisse pmd_t pmd) 52253f5c3f4SJérôme Glisse { 523992de9a8SJérôme Glisse #ifdef CONFIG_TRANSPARENT_HUGEPAGE 52453f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 525f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 5262aee09d8SJérôme Glisse unsigned long pfn, npages, i; 5272aee09d8SJérôme Glisse bool fault, write_fault; 528f88a1e90SJérôme Glisse uint64_t cpu_flags; 52953f5c3f4SJérôme Glisse 5302aee09d8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 531f88a1e90SJérôme Glisse cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); 5322aee09d8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, 5332aee09d8SJérôme Glisse &fault, &write_fault); 53453f5c3f4SJérôme Glisse 5352aee09d8SJérôme Glisse if (pmd_protnone(pmd) || fault || write_fault) 5362aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 53753f5c3f4SJérôme Glisse 53853f5c3f4SJérôme Glisse pfn = pmd_pfn(pmd) + pte_index(addr); 539992de9a8SJérôme Glisse for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) { 540992de9a8SJérôme Glisse if (pmd_devmap(pmd)) { 541992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 542992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 543992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 544992de9a8SJérôme Glisse return -EBUSY; 545992de9a8SJérôme Glisse } 546391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags; 547992de9a8SJérôme Glisse } 548992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 549992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 550992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 551992de9a8SJérôme Glisse } 55253f5c3f4SJérôme Glisse hmm_vma_walk->last = end; 55353f5c3f4SJérôme Glisse return 0; 554992de9a8SJérôme Glisse #else 555992de9a8SJérôme Glisse /* If THP is not enabled then we should never reach that code ! */ 556992de9a8SJérôme Glisse return -EINVAL; 557992de9a8SJérôme Glisse #endif 55853f5c3f4SJérôme Glisse } 55953f5c3f4SJérôme Glisse 560f88a1e90SJérôme Glisse static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) 5612aee09d8SJérôme Glisse { 562789c2af8SPhilip Yang if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte)) 5632aee09d8SJérôme Glisse return 0; 564f88a1e90SJérôme Glisse return pte_write(pte) ? range->flags[HMM_PFN_VALID] | 565f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 566f88a1e90SJérôme Glisse range->flags[HMM_PFN_VALID]; 5672aee09d8SJérôme Glisse } 5682aee09d8SJérôme Glisse 56953f5c3f4SJérôme Glisse static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, 57053f5c3f4SJérôme Glisse unsigned long end, pmd_t *pmdp, pte_t *ptep, 57153f5c3f4SJérôme Glisse uint64_t *pfn) 57253f5c3f4SJérôme Glisse { 57353f5c3f4SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 574f88a1e90SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 57553f5c3f4SJérôme Glisse struct vm_area_struct *vma = walk->vma; 5762aee09d8SJérôme Glisse bool fault, write_fault; 5772aee09d8SJérôme Glisse uint64_t cpu_flags; 57853f5c3f4SJérôme Glisse pte_t pte = *ptep; 579f88a1e90SJérôme Glisse uint64_t orig_pfn = *pfn; 58053f5c3f4SJérôme Glisse 581f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_NONE]; 58273231612SJérôme Glisse fault = write_fault = false; 58353f5c3f4SJérôme Glisse 58453f5c3f4SJérôme Glisse if (pte_none(pte)) { 58573231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0, 58673231612SJérôme Glisse &fault, &write_fault); 5872aee09d8SJérôme Glisse if (fault || write_fault) 58853f5c3f4SJérôme Glisse goto fault; 58953f5c3f4SJérôme Glisse return 0; 59053f5c3f4SJérôme Glisse } 59153f5c3f4SJérôme Glisse 59253f5c3f4SJérôme Glisse if (!pte_present(pte)) { 59353f5c3f4SJérôme Glisse swp_entry_t entry = pte_to_swp_entry(pte); 59453f5c3f4SJérôme Glisse 59553f5c3f4SJérôme Glisse if (!non_swap_entry(entry)) { 5962aee09d8SJérôme Glisse if (fault || write_fault) 59753f5c3f4SJérôme Glisse goto fault; 59853f5c3f4SJérôme Glisse return 0; 59953f5c3f4SJérôme Glisse } 60053f5c3f4SJérôme Glisse 60153f5c3f4SJérôme Glisse /* 60253f5c3f4SJérôme Glisse * This is a special swap entry, ignore migration, use 60353f5c3f4SJérôme Glisse * device and report anything else as error. 60453f5c3f4SJérôme Glisse */ 60553f5c3f4SJérôme Glisse if (is_device_private_entry(entry)) { 606f88a1e90SJérôme Glisse cpu_flags = range->flags[HMM_PFN_VALID] | 607f88a1e90SJérôme Glisse range->flags[HMM_PFN_DEVICE_PRIVATE]; 6082aee09d8SJérôme Glisse cpu_flags |= is_write_device_private_entry(entry) ? 609f88a1e90SJérôme Glisse range->flags[HMM_PFN_WRITE] : 0; 610f88a1e90SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 611f88a1e90SJérôme Glisse &fault, &write_fault); 612f88a1e90SJérôme Glisse if (fault || write_fault) 613f88a1e90SJérôme Glisse goto fault; 614391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, 615391aab11SJérôme Glisse swp_offset(entry)); 616f88a1e90SJérôme Glisse *pfn |= cpu_flags; 61753f5c3f4SJérôme Glisse return 0; 61853f5c3f4SJérôme Glisse } 61953f5c3f4SJérôme Glisse 62053f5c3f4SJérôme Glisse if (is_migration_entry(entry)) { 6212aee09d8SJérôme Glisse if (fault || write_fault) { 62253f5c3f4SJérôme Glisse pte_unmap(ptep); 62353f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 62453f5c3f4SJérôme Glisse migration_entry_wait(vma->vm_mm, 62553f5c3f4SJérôme Glisse pmdp, addr); 62673231612SJérôme Glisse return -EBUSY; 62753f5c3f4SJérôme Glisse } 62853f5c3f4SJérôme Glisse return 0; 62953f5c3f4SJérôme Glisse } 63053f5c3f4SJérôme Glisse 63153f5c3f4SJérôme Glisse /* Report error for everything else */ 632f88a1e90SJérôme Glisse *pfn = range->values[HMM_PFN_ERROR]; 63353f5c3f4SJérôme Glisse return -EFAULT; 63473231612SJérôme Glisse } else { 63573231612SJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, pte); 63673231612SJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 63773231612SJérôme Glisse &fault, &write_fault); 63853f5c3f4SJérôme Glisse } 63953f5c3f4SJérôme Glisse 6402aee09d8SJérôme Glisse if (fault || write_fault) 64153f5c3f4SJérôme Glisse goto fault; 64253f5c3f4SJérôme Glisse 643992de9a8SJérôme Glisse if (pte_devmap(pte)) { 644992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte), 645992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 646992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 647992de9a8SJérôme Glisse return -EBUSY; 648992de9a8SJérôme Glisse } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) { 649992de9a8SJérôme Glisse *pfn = range->values[HMM_PFN_SPECIAL]; 650992de9a8SJérôme Glisse return -EFAULT; 651992de9a8SJérôme Glisse } 652992de9a8SJérôme Glisse 653391aab11SJérôme Glisse *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags; 65453f5c3f4SJérôme Glisse return 0; 65553f5c3f4SJérôme Glisse 65653f5c3f4SJérôme Glisse fault: 657992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 658992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 659992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 660992de9a8SJérôme Glisse } 66153f5c3f4SJérôme Glisse pte_unmap(ptep); 66253f5c3f4SJérôme Glisse /* Fault any virtual address we were asked to fault */ 6632aee09d8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 66453f5c3f4SJérôme Glisse } 66553f5c3f4SJérôme Glisse 666da4c3c73SJérôme Glisse static int hmm_vma_walk_pmd(pmd_t *pmdp, 667da4c3c73SJérôme Glisse unsigned long start, 668da4c3c73SJérôme Glisse unsigned long end, 669da4c3c73SJérôme Glisse struct mm_walk *walk) 670da4c3c73SJérôme Glisse { 67174eee180SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 67274eee180SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 673d08faca0SJérôme Glisse struct vm_area_struct *vma = walk->vma; 674ff05c0c6SJérôme Glisse uint64_t *pfns = range->pfns; 675da4c3c73SJérôme Glisse unsigned long addr = start, i; 676da4c3c73SJérôme Glisse pte_t *ptep; 677da4c3c73SJérôme Glisse pmd_t pmd; 678da4c3c73SJérôme Glisse 679d08faca0SJérôme Glisse 680d08faca0SJérôme Glisse again: 681d08faca0SJérôme Glisse pmd = READ_ONCE(*pmdp); 682d08faca0SJérôme Glisse if (pmd_none(pmd)) 683d08faca0SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 684d08faca0SJérôme Glisse 685d08faca0SJérôme Glisse if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB)) 686d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 687d08faca0SJérôme Glisse 688d08faca0SJérôme Glisse if (thp_migration_supported() && is_pmd_migration_entry(pmd)) { 689d08faca0SJérôme Glisse bool fault, write_fault; 690d08faca0SJérôme Glisse unsigned long npages; 691d08faca0SJérôme Glisse uint64_t *pfns; 692d08faca0SJérôme Glisse 693d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 694d08faca0SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 695d08faca0SJérôme Glisse pfns = &range->pfns[i]; 696d08faca0SJérôme Glisse 697d08faca0SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 698d08faca0SJérôme Glisse 0, &fault, &write_fault); 699d08faca0SJérôme Glisse if (fault || write_fault) { 700d08faca0SJérôme Glisse hmm_vma_walk->last = addr; 701d08faca0SJérôme Glisse pmd_migration_entry_wait(vma->vm_mm, pmdp); 70273231612SJérôme Glisse return -EBUSY; 703d08faca0SJérôme Glisse } 704d08faca0SJérôme Glisse return 0; 705d08faca0SJérôme Glisse } else if (!pmd_present(pmd)) 706d08faca0SJérôme Glisse return hmm_pfns_bad(start, end, walk); 707d08faca0SJérôme Glisse 708d08faca0SJérôme Glisse if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) { 709da4c3c73SJérôme Glisse /* 710da4c3c73SJérôme Glisse * No need to take pmd_lock here, even if some other threads 711da4c3c73SJérôme Glisse * is splitting the huge pmd we will get that event through 712da4c3c73SJérôme Glisse * mmu_notifier callback. 713da4c3c73SJérôme Glisse * 714da4c3c73SJérôme Glisse * So just read pmd value and check again its a transparent 715da4c3c73SJérôme Glisse * huge or device mapping one and compute corresponding pfn 716da4c3c73SJérôme Glisse * values. 717da4c3c73SJérôme Glisse */ 718da4c3c73SJérôme Glisse pmd = pmd_read_atomic(pmdp); 719da4c3c73SJérôme Glisse barrier(); 720da4c3c73SJérôme Glisse if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) 721da4c3c73SJérôme Glisse goto again; 722da4c3c73SJérôme Glisse 723d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 72453f5c3f4SJérôme Glisse return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); 725da4c3c73SJérôme Glisse } 726da4c3c73SJérôme Glisse 727d08faca0SJérôme Glisse /* 728d08faca0SJérôme Glisse * We have handled all the valid case above ie either none, migration, 729d08faca0SJérôme Glisse * huge or transparent huge. At this point either it is a valid pmd 730d08faca0SJérôme Glisse * entry pointing to pte directory or it is a bad pmd that will not 731d08faca0SJérôme Glisse * recover. 732d08faca0SJérôme Glisse */ 733d08faca0SJérôme Glisse if (pmd_bad(pmd)) 734da4c3c73SJérôme Glisse return hmm_pfns_bad(start, end, walk); 735da4c3c73SJérôme Glisse 736da4c3c73SJérôme Glisse ptep = pte_offset_map(pmdp, addr); 737d08faca0SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 738da4c3c73SJérôme Glisse for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { 73953f5c3f4SJérôme Glisse int r; 740da4c3c73SJérôme Glisse 74153f5c3f4SJérôme Glisse r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); 74253f5c3f4SJérôme Glisse if (r) { 74353f5c3f4SJérôme Glisse /* hmm_vma_handle_pte() did unmap pte directory */ 74474eee180SJérôme Glisse hmm_vma_walk->last = addr; 74553f5c3f4SJérôme Glisse return r; 74674eee180SJérôme Glisse } 747da4c3c73SJérôme Glisse } 748992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 749992de9a8SJérôme Glisse /* 750992de9a8SJérôme Glisse * We do put_dev_pagemap() here and not in hmm_vma_handle_pte() 751992de9a8SJérôme Glisse * so that we can leverage get_dev_pagemap() optimization which 752992de9a8SJérôme Glisse * will not re-take a reference on a pgmap if we already have 753992de9a8SJérôme Glisse * one. 754992de9a8SJérôme Glisse */ 755992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 756992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 757992de9a8SJérôme Glisse } 758da4c3c73SJérôme Glisse pte_unmap(ptep - 1); 759da4c3c73SJérôme Glisse 76053f5c3f4SJérôme Glisse hmm_vma_walk->last = addr; 761da4c3c73SJérôme Glisse return 0; 762da4c3c73SJérôme Glisse } 763da4c3c73SJérôme Glisse 764992de9a8SJérôme Glisse static int hmm_vma_walk_pud(pud_t *pudp, 765992de9a8SJérôme Glisse unsigned long start, 766992de9a8SJérôme Glisse unsigned long end, 767992de9a8SJérôme Glisse struct mm_walk *walk) 768992de9a8SJérôme Glisse { 769992de9a8SJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 770992de9a8SJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 771992de9a8SJérôme Glisse unsigned long addr = start, next; 772992de9a8SJérôme Glisse pmd_t *pmdp; 773992de9a8SJérôme Glisse pud_t pud; 774992de9a8SJérôme Glisse int ret; 775992de9a8SJérôme Glisse 776992de9a8SJérôme Glisse again: 777992de9a8SJérôme Glisse pud = READ_ONCE(*pudp); 778992de9a8SJérôme Glisse if (pud_none(pud)) 779992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 780992de9a8SJérôme Glisse 781992de9a8SJérôme Glisse if (pud_huge(pud) && pud_devmap(pud)) { 782992de9a8SJérôme Glisse unsigned long i, npages, pfn; 783992de9a8SJérôme Glisse uint64_t *pfns, cpu_flags; 784992de9a8SJérôme Glisse bool fault, write_fault; 785992de9a8SJérôme Glisse 786992de9a8SJérôme Glisse if (!pud_present(pud)) 787992de9a8SJérôme Glisse return hmm_vma_walk_hole(start, end, walk); 788992de9a8SJérôme Glisse 789992de9a8SJérôme Glisse i = (addr - range->start) >> PAGE_SHIFT; 790992de9a8SJérôme Glisse npages = (end - addr) >> PAGE_SHIFT; 791992de9a8SJérôme Glisse pfns = &range->pfns[i]; 792992de9a8SJérôme Glisse 793992de9a8SJérôme Glisse cpu_flags = pud_to_hmm_pfn_flags(range, pud); 794992de9a8SJérôme Glisse hmm_range_need_fault(hmm_vma_walk, pfns, npages, 795992de9a8SJérôme Glisse cpu_flags, &fault, &write_fault); 796992de9a8SJérôme Glisse if (fault || write_fault) 797992de9a8SJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, 798992de9a8SJérôme Glisse write_fault, walk); 799992de9a8SJérôme Glisse 800992de9a8SJérôme Glisse pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); 801992de9a8SJérôme Glisse for (i = 0; i < npages; ++i, ++pfn) { 802992de9a8SJérôme Glisse hmm_vma_walk->pgmap = get_dev_pagemap(pfn, 803992de9a8SJérôme Glisse hmm_vma_walk->pgmap); 804992de9a8SJérôme Glisse if (unlikely(!hmm_vma_walk->pgmap)) 805992de9a8SJérôme Glisse return -EBUSY; 806391aab11SJérôme Glisse pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 807391aab11SJérôme Glisse cpu_flags; 808992de9a8SJérôme Glisse } 809992de9a8SJérôme Glisse if (hmm_vma_walk->pgmap) { 810992de9a8SJérôme Glisse put_dev_pagemap(hmm_vma_walk->pgmap); 811992de9a8SJérôme Glisse hmm_vma_walk->pgmap = NULL; 812992de9a8SJérôme Glisse } 813992de9a8SJérôme Glisse hmm_vma_walk->last = end; 814992de9a8SJérôme Glisse return 0; 815992de9a8SJérôme Glisse } 816992de9a8SJérôme Glisse 817992de9a8SJérôme Glisse split_huge_pud(walk->vma, pudp, addr); 818992de9a8SJérôme Glisse if (pud_none(*pudp)) 819992de9a8SJérôme Glisse goto again; 820992de9a8SJérôme Glisse 821992de9a8SJérôme Glisse pmdp = pmd_offset(pudp, addr); 822992de9a8SJérôme Glisse do { 823992de9a8SJérôme Glisse next = pmd_addr_end(addr, end); 824992de9a8SJérôme Glisse ret = hmm_vma_walk_pmd(pmdp, addr, next, walk); 825992de9a8SJérôme Glisse if (ret) 826992de9a8SJérôme Glisse return ret; 827992de9a8SJérôme Glisse } while (pmdp++, addr = next, addr != end); 828992de9a8SJérôme Glisse 829992de9a8SJérôme Glisse return 0; 830992de9a8SJérôme Glisse } 831992de9a8SJérôme Glisse 83263d5066fSJérôme Glisse static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask, 83363d5066fSJérôme Glisse unsigned long start, unsigned long end, 83463d5066fSJérôme Glisse struct mm_walk *walk) 83563d5066fSJérôme Glisse { 83663d5066fSJérôme Glisse #ifdef CONFIG_HUGETLB_PAGE 83763d5066fSJérôme Glisse unsigned long addr = start, i, pfn, mask, size, pfn_inc; 83863d5066fSJérôme Glisse struct hmm_vma_walk *hmm_vma_walk = walk->private; 83963d5066fSJérôme Glisse struct hmm_range *range = hmm_vma_walk->range; 84063d5066fSJérôme Glisse struct vm_area_struct *vma = walk->vma; 84163d5066fSJérôme Glisse struct hstate *h = hstate_vma(vma); 84263d5066fSJérôme Glisse uint64_t orig_pfn, cpu_flags; 84363d5066fSJérôme Glisse bool fault, write_fault; 84463d5066fSJérôme Glisse spinlock_t *ptl; 84563d5066fSJérôme Glisse pte_t entry; 84663d5066fSJérôme Glisse int ret = 0; 84763d5066fSJérôme Glisse 84863d5066fSJérôme Glisse size = 1UL << huge_page_shift(h); 84963d5066fSJérôme Glisse mask = size - 1; 85063d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) { 85163d5066fSJérôme Glisse /* Make sure we are looking at full page. */ 85263d5066fSJérôme Glisse if (start & mask) 85363d5066fSJérôme Glisse return -EINVAL; 85463d5066fSJérôme Glisse if (end < (start + size)) 85563d5066fSJérôme Glisse return -EINVAL; 85663d5066fSJérôme Glisse pfn_inc = size >> PAGE_SHIFT; 85763d5066fSJérôme Glisse } else { 85863d5066fSJérôme Glisse pfn_inc = 1; 85963d5066fSJérôme Glisse size = PAGE_SIZE; 86063d5066fSJérôme Glisse } 86163d5066fSJérôme Glisse 86263d5066fSJérôme Glisse 86363d5066fSJérôme Glisse ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); 86463d5066fSJérôme Glisse entry = huge_ptep_get(pte); 86563d5066fSJérôme Glisse 86663d5066fSJérôme Glisse i = (start - range->start) >> range->page_shift; 86763d5066fSJérôme Glisse orig_pfn = range->pfns[i]; 86863d5066fSJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 86963d5066fSJérôme Glisse cpu_flags = pte_to_hmm_pfn_flags(range, entry); 87063d5066fSJérôme Glisse fault = write_fault = false; 87163d5066fSJérôme Glisse hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, 87263d5066fSJérôme Glisse &fault, &write_fault); 87363d5066fSJérôme Glisse if (fault || write_fault) { 87463d5066fSJérôme Glisse ret = -ENOENT; 87563d5066fSJérôme Glisse goto unlock; 87663d5066fSJérôme Glisse } 87763d5066fSJérôme Glisse 87863d5066fSJérôme Glisse pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift); 87963d5066fSJérôme Glisse for (; addr < end; addr += size, i++, pfn += pfn_inc) 880391aab11SJérôme Glisse range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) | 881391aab11SJérôme Glisse cpu_flags; 88263d5066fSJérôme Glisse hmm_vma_walk->last = end; 88363d5066fSJérôme Glisse 88463d5066fSJérôme Glisse unlock: 88563d5066fSJérôme Glisse spin_unlock(ptl); 88663d5066fSJérôme Glisse 88763d5066fSJérôme Glisse if (ret == -ENOENT) 88863d5066fSJérôme Glisse return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); 88963d5066fSJérôme Glisse 89063d5066fSJérôme Glisse return ret; 89163d5066fSJérôme Glisse #else /* CONFIG_HUGETLB_PAGE */ 89263d5066fSJérôme Glisse return -EINVAL; 89363d5066fSJérôme Glisse #endif 89463d5066fSJérôme Glisse } 89563d5066fSJérôme Glisse 896f88a1e90SJérôme Glisse static void hmm_pfns_clear(struct hmm_range *range, 897f88a1e90SJérôme Glisse uint64_t *pfns, 89833cd47dcSJérôme Glisse unsigned long addr, 89933cd47dcSJérôme Glisse unsigned long end) 90033cd47dcSJérôme Glisse { 90133cd47dcSJérôme Glisse for (; addr < end; addr += PAGE_SIZE, pfns++) 902f88a1e90SJérôme Glisse *pfns = range->values[HMM_PFN_NONE]; 90333cd47dcSJérôme Glisse } 90433cd47dcSJérôme Glisse 905da4c3c73SJérôme Glisse /* 906a3e0d41cSJérôme Glisse * hmm_range_register() - start tracking change to CPU page table over a range 907a3e0d41cSJérôme Glisse * @range: range 908a3e0d41cSJérôme Glisse * @mm: the mm struct for the range of virtual address 909a3e0d41cSJérôme Glisse * @start: start virtual address (inclusive) 910a3e0d41cSJérôme Glisse * @end: end virtual address (exclusive) 91163d5066fSJérôme Glisse * @page_shift: expect page shift for the range 912a3e0d41cSJérôme Glisse * Returns 0 on success, -EFAULT if the address space is no longer valid 913a3e0d41cSJérôme Glisse * 914a3e0d41cSJérôme Glisse * Track updates to the CPU page table see include/linux/hmm.h 915a3e0d41cSJérôme Glisse */ 916a3e0d41cSJérôme Glisse int hmm_range_register(struct hmm_range *range, 917*e36acfe6SJason Gunthorpe struct hmm_mirror *mirror, 918a3e0d41cSJérôme Glisse unsigned long start, 91963d5066fSJérôme Glisse unsigned long end, 92063d5066fSJérôme Glisse unsigned page_shift) 921a3e0d41cSJérôme Glisse { 92263d5066fSJérôme Glisse unsigned long mask = ((1UL << page_shift) - 1UL); 923*e36acfe6SJason Gunthorpe struct hmm *hmm = mirror->hmm; 92463d5066fSJérôme Glisse 925a3e0d41cSJérôme Glisse range->valid = false; 926a3e0d41cSJérôme Glisse range->hmm = NULL; 927a3e0d41cSJérôme Glisse 92863d5066fSJérôme Glisse if ((start & mask) || (end & mask)) 92963d5066fSJérôme Glisse return -EINVAL; 93063d5066fSJérôme Glisse if (start >= end) 931a3e0d41cSJérôme Glisse return -EINVAL; 932a3e0d41cSJérôme Glisse 93363d5066fSJérôme Glisse range->page_shift = page_shift; 934a3e0d41cSJérôme Glisse range->start = start; 935a3e0d41cSJérôme Glisse range->end = end; 936a3e0d41cSJérôme Glisse 937a3e0d41cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 938*e36acfe6SJason Gunthorpe if (hmm->mm == NULL || hmm->dead) 939a3e0d41cSJérôme Glisse return -EFAULT; 940a3e0d41cSJérôme Glisse 941085ea250SRalph Campbell /* Initialize range to track CPU page table updates. */ 942085ea250SRalph Campbell mutex_lock(&hmm->lock); 943a3e0d41cSJérôme Glisse 944085ea250SRalph Campbell range->hmm = hmm; 945*e36acfe6SJason Gunthorpe kref_get(&hmm->kref); 946085ea250SRalph Campbell list_add_rcu(&range->list, &hmm->ranges); 947a3e0d41cSJérôme Glisse 948a3e0d41cSJérôme Glisse /* 949a3e0d41cSJérôme Glisse * If there are any concurrent notifiers we have to wait for them for 950a3e0d41cSJérôme Glisse * the range to be valid (see hmm_range_wait_until_valid()). 951a3e0d41cSJérôme Glisse */ 952085ea250SRalph Campbell if (!hmm->notifiers) 953a3e0d41cSJérôme Glisse range->valid = true; 954085ea250SRalph Campbell mutex_unlock(&hmm->lock); 955a3e0d41cSJérôme Glisse 956a3e0d41cSJérôme Glisse return 0; 957a3e0d41cSJérôme Glisse } 958a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_register); 959a3e0d41cSJérôme Glisse 960a3e0d41cSJérôme Glisse /* 961a3e0d41cSJérôme Glisse * hmm_range_unregister() - stop tracking change to CPU page table over a range 962a3e0d41cSJérôme Glisse * @range: range 963a3e0d41cSJérôme Glisse * 964a3e0d41cSJérôme Glisse * Range struct is used to track updates to the CPU page table after a call to 965a3e0d41cSJérôme Glisse * hmm_range_register(). See include/linux/hmm.h for how to use it. 966a3e0d41cSJérôme Glisse */ 967a3e0d41cSJérôme Glisse void hmm_range_unregister(struct hmm_range *range) 968a3e0d41cSJérôme Glisse { 969085ea250SRalph Campbell struct hmm *hmm = range->hmm; 970085ea250SRalph Campbell 971a3e0d41cSJérôme Glisse /* Sanity check this really should not happen. */ 972085ea250SRalph Campbell if (hmm == NULL || range->end <= range->start) 973a3e0d41cSJérôme Glisse return; 974a3e0d41cSJérôme Glisse 975085ea250SRalph Campbell mutex_lock(&hmm->lock); 976a3e0d41cSJérôme Glisse list_del_rcu(&range->list); 977085ea250SRalph Campbell mutex_unlock(&hmm->lock); 978a3e0d41cSJérôme Glisse 979a3e0d41cSJérôme Glisse /* Drop reference taken by hmm_range_register() */ 980a3e0d41cSJérôme Glisse range->valid = false; 981085ea250SRalph Campbell hmm_put(hmm); 982a3e0d41cSJérôme Glisse range->hmm = NULL; 983a3e0d41cSJérôme Glisse } 984a3e0d41cSJérôme Glisse EXPORT_SYMBOL(hmm_range_unregister); 985a3e0d41cSJérôme Glisse 986a3e0d41cSJérôme Glisse /* 98725f23a0cSJérôme Glisse * hmm_range_snapshot() - snapshot CPU page table for a range 98825f23a0cSJérôme Glisse * @range: range 989085ea250SRalph Campbell * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid 990a3e0d41cSJérôme Glisse * permission (for instance asking for write and range is read only), 991a3e0d41cSJérôme Glisse * -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid 992a3e0d41cSJérôme Glisse * vma or it is illegal to access that range), number of valid pages 993a3e0d41cSJérôme Glisse * in range->pfns[] (from range start address). 994da4c3c73SJérôme Glisse * 995da4c3c73SJérôme Glisse * This snapshots the CPU page table for a range of virtual addresses. Snapshot 996a3e0d41cSJérôme Glisse * validity is tracked by range struct. See in include/linux/hmm.h for example 997a3e0d41cSJérôme Glisse * on how to use. 998da4c3c73SJérôme Glisse */ 99925f23a0cSJérôme Glisse long hmm_range_snapshot(struct hmm_range *range) 1000da4c3c73SJérôme Glisse { 100163d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1002a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 100374eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1004a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1005a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 1006da4c3c73SJérôme Glisse struct mm_walk mm_walk; 1007704f3f2cSJérôme Glisse 1008704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1009a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1010a3e0d41cSJérôme Glisse return -EFAULT; 1011da4c3c73SJérôme Glisse 1012a3e0d41cSJérôme Glisse do { 1013a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1014a3e0d41cSJérôme Glisse if (!range->valid) 1015a3e0d41cSJérôme Glisse return -EAGAIN; 1016a3e0d41cSJérôme Glisse 1017a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 101863d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1019a3e0d41cSJérôme Glisse return -EFAULT; 1020a3e0d41cSJérôme Glisse 102163d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 10221c2308f0SJason Gunthorpe if (huge_page_shift(hstate_vma(vma)) != 10231c2308f0SJason Gunthorpe range->page_shift && 102463d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 102563d5066fSJérôme Glisse return -EINVAL; 102663d5066fSJérôme Glisse } else { 102763d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 102863d5066fSJérôme Glisse return -EINVAL; 102963d5066fSJérôme Glisse } 103063d5066fSJérôme Glisse 103186586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 103286586a41SJérôme Glisse /* 1033a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1034a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1035a3e0d41cSJérôme Glisse * support architecture that allow write without read. 103686586a41SJérôme Glisse */ 1037a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1038a3e0d41cSJérôme Glisse range->start, range->end); 103986586a41SJérôme Glisse return -EPERM; 104086586a41SJérôme Glisse } 104186586a41SJérôme Glisse 1042a3e0d41cSJérôme Glisse range->vma = vma; 1043992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1044a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 104574eee180SJérôme Glisse hmm_vma_walk.fault = false; 104674eee180SJérôme Glisse hmm_vma_walk.range = range; 104774eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1048a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 104974eee180SJérôme Glisse 1050da4c3c73SJérôme Glisse mm_walk.vma = vma; 1051da4c3c73SJérôme Glisse mm_walk.mm = vma->vm_mm; 1052da4c3c73SJérôme Glisse mm_walk.pte_entry = NULL; 1053da4c3c73SJérôme Glisse mm_walk.test_walk = NULL; 1054da4c3c73SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1055992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 1056da4c3c73SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 1057da4c3c73SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 105863d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 1059da4c3c73SJérôme Glisse 1060a3e0d41cSJérôme Glisse walk_page_range(start, end, &mm_walk); 1061a3e0d41cSJérôme Glisse start = end; 1062a3e0d41cSJérôme Glisse } while (start < range->end); 1063a3e0d41cSJérôme Glisse 106425f23a0cSJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1065da4c3c73SJérôme Glisse } 106625f23a0cSJérôme Glisse EXPORT_SYMBOL(hmm_range_snapshot); 1067da4c3c73SJérôme Glisse 1068da4c3c73SJérôme Glisse /* 106973231612SJérôme Glisse * hmm_range_fault() - try to fault some address in a virtual address range 107008232a45SJérôme Glisse * @range: range being faulted 107174eee180SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1072085ea250SRalph Campbell * Return: number of valid pages in range->pfns[] (from range start 107373231612SJérôme Glisse * address). This may be zero. If the return value is negative, 107473231612SJérôme Glisse * then one of the following values may be returned: 107573231612SJérôme Glisse * 107673231612SJérôme Glisse * -EINVAL invalid arguments or mm or virtual address are in an 107763d5066fSJérôme Glisse * invalid vma (for instance device file vma). 107873231612SJérôme Glisse * -ENOMEM: Out of memory. 107973231612SJérôme Glisse * -EPERM: Invalid permission (for instance asking for write and 108073231612SJérôme Glisse * range is read only). 108173231612SJérôme Glisse * -EAGAIN: If you need to retry and mmap_sem was drop. This can only 108273231612SJérôme Glisse * happens if block argument is false. 108373231612SJérôme Glisse * -EBUSY: If the the range is being invalidated and you should wait 108473231612SJérôme Glisse * for invalidation to finish. 108573231612SJérôme Glisse * -EFAULT: Invalid (ie either no valid vma or it is illegal to access 108673231612SJérôme Glisse * that range), number of valid pages in range->pfns[] (from 108773231612SJérôme Glisse * range start address). 108874eee180SJérôme Glisse * 108974eee180SJérôme Glisse * This is similar to a regular CPU page fault except that it will not trigger 109073231612SJérôme Glisse * any memory migration if the memory being faulted is not accessible by CPUs 109173231612SJérôme Glisse * and caller does not ask for migration. 109274eee180SJérôme Glisse * 1093ff05c0c6SJérôme Glisse * On error, for one virtual address in the range, the function will mark the 1094ff05c0c6SJérôme Glisse * corresponding HMM pfn entry with an error flag. 109574eee180SJérôme Glisse */ 109673231612SJérôme Glisse long hmm_range_fault(struct hmm_range *range, bool block) 109774eee180SJérôme Glisse { 109863d5066fSJérôme Glisse const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 1099a3e0d41cSJérôme Glisse unsigned long start = range->start, end; 110074eee180SJérôme Glisse struct hmm_vma_walk hmm_vma_walk; 1101a3e0d41cSJérôme Glisse struct hmm *hmm = range->hmm; 1102a3e0d41cSJérôme Glisse struct vm_area_struct *vma; 110374eee180SJérôme Glisse struct mm_walk mm_walk; 110474eee180SJérôme Glisse int ret; 110574eee180SJérôme Glisse 1106704f3f2cSJérôme Glisse /* Check if hmm_mm_destroy() was call. */ 1107a3e0d41cSJérôme Glisse if (hmm->mm == NULL || hmm->dead) 1108a3e0d41cSJérôme Glisse return -EFAULT; 1109a3e0d41cSJérôme Glisse 1110a3e0d41cSJérôme Glisse do { 1111a3e0d41cSJérôme Glisse /* If range is no longer valid force retry. */ 1112a3e0d41cSJérôme Glisse if (!range->valid) { 1113a3e0d41cSJérôme Glisse up_read(&hmm->mm->mmap_sem); 1114a3e0d41cSJérôme Glisse return -EAGAIN; 1115704f3f2cSJérôme Glisse } 111674eee180SJérôme Glisse 1117a3e0d41cSJérôme Glisse vma = find_vma(hmm->mm, start); 111863d5066fSJérôme Glisse if (vma == NULL || (vma->vm_flags & device_vma)) 1119a3e0d41cSJérôme Glisse return -EFAULT; 1120a3e0d41cSJérôme Glisse 112163d5066fSJérôme Glisse if (is_vm_hugetlb_page(vma)) { 112263d5066fSJérôme Glisse if (huge_page_shift(hstate_vma(vma)) != 112363d5066fSJérôme Glisse range->page_shift && 112463d5066fSJérôme Glisse range->page_shift != PAGE_SHIFT) 112563d5066fSJérôme Glisse return -EINVAL; 112663d5066fSJérôme Glisse } else { 112763d5066fSJérôme Glisse if (range->page_shift != PAGE_SHIFT) 112863d5066fSJérôme Glisse return -EINVAL; 112963d5066fSJérôme Glisse } 113063d5066fSJérôme Glisse 113186586a41SJérôme Glisse if (!(vma->vm_flags & VM_READ)) { 113286586a41SJérôme Glisse /* 1133a3e0d41cSJérôme Glisse * If vma do not allow read access, then assume that it 1134a3e0d41cSJérôme Glisse * does not allow write access, either. HMM does not 1135a3e0d41cSJérôme Glisse * support architecture that allow write without read. 113686586a41SJérôme Glisse */ 1137a3e0d41cSJérôme Glisse hmm_pfns_clear(range, range->pfns, 1138a3e0d41cSJérôme Glisse range->start, range->end); 113986586a41SJérôme Glisse return -EPERM; 114086586a41SJérôme Glisse } 114174eee180SJérôme Glisse 1142a3e0d41cSJérôme Glisse range->vma = vma; 1143992de9a8SJérôme Glisse hmm_vma_walk.pgmap = NULL; 1144a3e0d41cSJérôme Glisse hmm_vma_walk.last = start; 114574eee180SJérôme Glisse hmm_vma_walk.fault = true; 114674eee180SJérôme Glisse hmm_vma_walk.block = block; 114774eee180SJérôme Glisse hmm_vma_walk.range = range; 114874eee180SJérôme Glisse mm_walk.private = &hmm_vma_walk; 1149a3e0d41cSJérôme Glisse end = min(range->end, vma->vm_end); 115074eee180SJérôme Glisse 115174eee180SJérôme Glisse mm_walk.vma = vma; 115274eee180SJérôme Glisse mm_walk.mm = vma->vm_mm; 115374eee180SJérôme Glisse mm_walk.pte_entry = NULL; 115474eee180SJérôme Glisse mm_walk.test_walk = NULL; 115574eee180SJérôme Glisse mm_walk.hugetlb_entry = NULL; 1156992de9a8SJérôme Glisse mm_walk.pud_entry = hmm_vma_walk_pud; 115774eee180SJérôme Glisse mm_walk.pmd_entry = hmm_vma_walk_pmd; 115874eee180SJérôme Glisse mm_walk.pte_hole = hmm_vma_walk_hole; 115963d5066fSJérôme Glisse mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; 116074eee180SJérôme Glisse 116174eee180SJérôme Glisse do { 1162a3e0d41cSJérôme Glisse ret = walk_page_range(start, end, &mm_walk); 116374eee180SJérôme Glisse start = hmm_vma_walk.last; 1164a3e0d41cSJérôme Glisse 116573231612SJérôme Glisse /* Keep trying while the range is valid. */ 116673231612SJérôme Glisse } while (ret == -EBUSY && range->valid); 116774eee180SJérôme Glisse 116874eee180SJérôme Glisse if (ret) { 116974eee180SJérôme Glisse unsigned long i; 117074eee180SJérôme Glisse 117174eee180SJérôme Glisse i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 1172a3e0d41cSJérôme Glisse hmm_pfns_clear(range, &range->pfns[i], 1173a3e0d41cSJérôme Glisse hmm_vma_walk.last, range->end); 117473231612SJérôme Glisse return ret; 117574eee180SJérôme Glisse } 1176a3e0d41cSJérôme Glisse start = end; 1177a3e0d41cSJérôme Glisse 1178a3e0d41cSJérôme Glisse } while (start < range->end); 1179704f3f2cSJérôme Glisse 118073231612SJérôme Glisse return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; 118174eee180SJérôme Glisse } 118273231612SJérôme Glisse EXPORT_SYMBOL(hmm_range_fault); 118355c0ece8SJérôme Glisse 118455c0ece8SJérôme Glisse /** 118555c0ece8SJérôme Glisse * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one. 118655c0ece8SJérôme Glisse * @range: range being faulted 118755c0ece8SJérôme Glisse * @device: device against to dma map page to 118855c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 118955c0ece8SJérôme Glisse * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) 1190085ea250SRalph Campbell * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been 119155c0ece8SJérôme Glisse * drop and you need to try again, some other error value otherwise 119255c0ece8SJérôme Glisse * 119355c0ece8SJérôme Glisse * Note same usage pattern as hmm_range_fault(). 119455c0ece8SJérôme Glisse */ 119555c0ece8SJérôme Glisse long hmm_range_dma_map(struct hmm_range *range, 119655c0ece8SJérôme Glisse struct device *device, 119755c0ece8SJérôme Glisse dma_addr_t *daddrs, 119855c0ece8SJérôme Glisse bool block) 119955c0ece8SJérôme Glisse { 120055c0ece8SJérôme Glisse unsigned long i, npages, mapped; 120155c0ece8SJérôme Glisse long ret; 120255c0ece8SJérôme Glisse 120355c0ece8SJérôme Glisse ret = hmm_range_fault(range, block); 120455c0ece8SJérôme Glisse if (ret <= 0) 120555c0ece8SJérôme Glisse return ret ? ret : -EBUSY; 120655c0ece8SJérôme Glisse 120755c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 120855c0ece8SJérôme Glisse for (i = 0, mapped = 0; i < npages; ++i) { 120955c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 121055c0ece8SJérôme Glisse struct page *page; 121155c0ece8SJérôme Glisse 121255c0ece8SJérôme Glisse /* 121355c0ece8SJérôme Glisse * FIXME need to update DMA API to provide invalid DMA address 121455c0ece8SJérôme Glisse * value instead of a function to test dma address value. This 121555c0ece8SJérôme Glisse * would remove lot of dumb code duplicated accross many arch. 121655c0ece8SJérôme Glisse * 121755c0ece8SJérôme Glisse * For now setting it to 0 here is good enough as the pfns[] 121855c0ece8SJérôme Glisse * value is what is use to check what is valid and what isn't. 121955c0ece8SJérôme Glisse */ 122055c0ece8SJérôme Glisse daddrs[i] = 0; 122155c0ece8SJérôme Glisse 1222391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 122355c0ece8SJérôme Glisse if (page == NULL) 122455c0ece8SJérôme Glisse continue; 122555c0ece8SJérôme Glisse 122655c0ece8SJérôme Glisse /* Check if range is being invalidated */ 122755c0ece8SJérôme Glisse if (!range->valid) { 122855c0ece8SJérôme Glisse ret = -EBUSY; 122955c0ece8SJérôme Glisse goto unmap; 123055c0ece8SJérôme Glisse } 123155c0ece8SJérôme Glisse 123255c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 123355c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 123455c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 123555c0ece8SJérôme Glisse 123655c0ece8SJérôme Glisse daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir); 123755c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) { 123855c0ece8SJérôme Glisse ret = -EFAULT; 123955c0ece8SJérôme Glisse goto unmap; 124055c0ece8SJérôme Glisse } 124155c0ece8SJérôme Glisse 124255c0ece8SJérôme Glisse mapped++; 124355c0ece8SJérôme Glisse } 124455c0ece8SJérôme Glisse 124555c0ece8SJérôme Glisse return mapped; 124655c0ece8SJérôme Glisse 124755c0ece8SJérôme Glisse unmap: 124855c0ece8SJérôme Glisse for (npages = i, i = 0; (i < npages) && mapped; ++i) { 124955c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 125055c0ece8SJérôme Glisse struct page *page; 125155c0ece8SJérôme Glisse 1252391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 125355c0ece8SJérôme Glisse if (page == NULL) 125455c0ece8SJérôme Glisse continue; 125555c0ece8SJérôme Glisse 125655c0ece8SJérôme Glisse if (dma_mapping_error(device, daddrs[i])) 125755c0ece8SJérôme Glisse continue; 125855c0ece8SJérôme Glisse 125955c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 126055c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) 126155c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 126255c0ece8SJérôme Glisse 126355c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 126455c0ece8SJérôme Glisse mapped--; 126555c0ece8SJérôme Glisse } 126655c0ece8SJérôme Glisse 126755c0ece8SJérôme Glisse return ret; 126855c0ece8SJérôme Glisse } 126955c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_map); 127055c0ece8SJérôme Glisse 127155c0ece8SJérôme Glisse /** 127255c0ece8SJérôme Glisse * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map() 127355c0ece8SJérôme Glisse * @range: range being unmapped 127455c0ece8SJérôme Glisse * @vma: the vma against which the range (optional) 127555c0ece8SJérôme Glisse * @device: device against which dma map was done 127655c0ece8SJérôme Glisse * @daddrs: dma address of mapped pages 127755c0ece8SJérôme Glisse * @dirty: dirty page if it had the write flag set 1278085ea250SRalph Campbell * Return: number of page unmapped on success, -EINVAL otherwise 127955c0ece8SJérôme Glisse * 128055c0ece8SJérôme Glisse * Note that caller MUST abide by mmu notifier or use HMM mirror and abide 128155c0ece8SJérôme Glisse * to the sync_cpu_device_pagetables() callback so that it is safe here to 128255c0ece8SJérôme Glisse * call set_page_dirty(). Caller must also take appropriate locks to avoid 128355c0ece8SJérôme Glisse * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress. 128455c0ece8SJérôme Glisse */ 128555c0ece8SJérôme Glisse long hmm_range_dma_unmap(struct hmm_range *range, 128655c0ece8SJérôme Glisse struct vm_area_struct *vma, 128755c0ece8SJérôme Glisse struct device *device, 128855c0ece8SJérôme Glisse dma_addr_t *daddrs, 128955c0ece8SJérôme Glisse bool dirty) 129055c0ece8SJérôme Glisse { 129155c0ece8SJérôme Glisse unsigned long i, npages; 129255c0ece8SJérôme Glisse long cpages = 0; 129355c0ece8SJérôme Glisse 129455c0ece8SJérôme Glisse /* Sanity check. */ 129555c0ece8SJérôme Glisse if (range->end <= range->start) 129655c0ece8SJérôme Glisse return -EINVAL; 129755c0ece8SJérôme Glisse if (!daddrs) 129855c0ece8SJérôme Glisse return -EINVAL; 129955c0ece8SJérôme Glisse if (!range->pfns) 130055c0ece8SJérôme Glisse return -EINVAL; 130155c0ece8SJérôme Glisse 130255c0ece8SJérôme Glisse npages = (range->end - range->start) >> PAGE_SHIFT; 130355c0ece8SJérôme Glisse for (i = 0; i < npages; ++i) { 130455c0ece8SJérôme Glisse enum dma_data_direction dir = DMA_TO_DEVICE; 130555c0ece8SJérôme Glisse struct page *page; 130655c0ece8SJérôme Glisse 1307391aab11SJérôme Glisse page = hmm_device_entry_to_page(range, range->pfns[i]); 130855c0ece8SJérôme Glisse if (page == NULL) 130955c0ece8SJérôme Glisse continue; 131055c0ece8SJérôme Glisse 131155c0ece8SJérôme Glisse /* If it is read and write than map bi-directional. */ 131255c0ece8SJérôme Glisse if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) { 131355c0ece8SJérôme Glisse dir = DMA_BIDIRECTIONAL; 131455c0ece8SJérôme Glisse 131555c0ece8SJérôme Glisse /* 131655c0ece8SJérôme Glisse * See comments in function description on why it is 131755c0ece8SJérôme Glisse * safe here to call set_page_dirty() 131855c0ece8SJérôme Glisse */ 131955c0ece8SJérôme Glisse if (dirty) 132055c0ece8SJérôme Glisse set_page_dirty(page); 132155c0ece8SJérôme Glisse } 132255c0ece8SJérôme Glisse 132355c0ece8SJérôme Glisse /* Unmap and clear pfns/dma address */ 132455c0ece8SJérôme Glisse dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); 132555c0ece8SJérôme Glisse range->pfns[i] = range->values[HMM_PFN_NONE]; 132655c0ece8SJérôme Glisse /* FIXME see comments in hmm_vma_dma_map() */ 132755c0ece8SJérôme Glisse daddrs[i] = 0; 132855c0ece8SJérôme Glisse cpages++; 132955c0ece8SJérôme Glisse } 133055c0ece8SJérôme Glisse 133155c0ece8SJérôme Glisse return cpages; 133255c0ece8SJérôme Glisse } 133355c0ece8SJérôme Glisse EXPORT_SYMBOL(hmm_range_dma_unmap); 1334c0b12405SJérôme Glisse #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ 13354ef589dcSJérôme Glisse 13364ef589dcSJérôme Glisse 1337df6ad698SJérôme Glisse #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) 13384ef589dcSJérôme Glisse struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, 13394ef589dcSJérôme Glisse unsigned long addr) 13404ef589dcSJérôme Glisse { 13414ef589dcSJérôme Glisse struct page *page; 13424ef589dcSJérôme Glisse 13434ef589dcSJérôme Glisse page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 13444ef589dcSJérôme Glisse if (!page) 13454ef589dcSJérôme Glisse return NULL; 13464ef589dcSJérôme Glisse lock_page(page); 13474ef589dcSJérôme Glisse return page; 13484ef589dcSJérôme Glisse } 13494ef589dcSJérôme Glisse EXPORT_SYMBOL(hmm_vma_alloc_locked_page); 13504ef589dcSJérôme Glisse 13514ef589dcSJérôme Glisse 13524ef589dcSJérôme Glisse static void hmm_devmem_ref_release(struct percpu_ref *ref) 13534ef589dcSJérôme Glisse { 13544ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13554ef589dcSJérôme Glisse 13564ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 13574ef589dcSJérôme Glisse complete(&devmem->completion); 13584ef589dcSJérôme Glisse } 13594ef589dcSJérôme Glisse 13604ef589dcSJérôme Glisse static void hmm_devmem_ref_exit(void *data) 13614ef589dcSJérôme Glisse { 13624ef589dcSJérôme Glisse struct percpu_ref *ref = data; 13634ef589dcSJérôme Glisse struct hmm_devmem *devmem; 13644ef589dcSJérôme Glisse 13654ef589dcSJérôme Glisse devmem = container_of(ref, struct hmm_devmem, ref); 1366bbecd94eSDan Williams wait_for_completion(&devmem->completion); 13674ef589dcSJérôme Glisse percpu_ref_exit(ref); 13684ef589dcSJérôme Glisse } 13694ef589dcSJérôme Glisse 1370bbecd94eSDan Williams static void hmm_devmem_ref_kill(struct percpu_ref *ref) 13714ef589dcSJérôme Glisse { 13724ef589dcSJérôme Glisse percpu_ref_kill(ref); 13734ef589dcSJérôme Glisse } 13744ef589dcSJérôme Glisse 1375b57e622eSSouptick Joarder static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma, 13764ef589dcSJérôme Glisse unsigned long addr, 13774ef589dcSJérôme Glisse const struct page *page, 13784ef589dcSJérôme Glisse unsigned int flags, 13794ef589dcSJérôme Glisse pmd_t *pmdp) 13804ef589dcSJérôme Glisse { 13814ef589dcSJérôme Glisse struct hmm_devmem *devmem = page->pgmap->data; 13824ef589dcSJérôme Glisse 13834ef589dcSJérôme Glisse return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); 13844ef589dcSJérôme Glisse } 13854ef589dcSJérôme Glisse 13864ef589dcSJérôme Glisse static void hmm_devmem_free(struct page *page, void *data) 13874ef589dcSJérôme Glisse { 13884ef589dcSJérôme Glisse struct hmm_devmem *devmem = data; 13894ef589dcSJérôme Glisse 13902fa147bdSDan Williams page->mapping = NULL; 13912fa147bdSDan Williams 13924ef589dcSJérôme Glisse devmem->ops->free(devmem, page); 13934ef589dcSJérôme Glisse } 13944ef589dcSJérôme Glisse 13954ef589dcSJérôme Glisse /* 13964ef589dcSJérôme Glisse * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory 13974ef589dcSJérôme Glisse * 13984ef589dcSJérôme Glisse * @ops: memory event device driver callback (see struct hmm_devmem_ops) 13994ef589dcSJérôme Glisse * @device: device struct to bind the resource too 14004ef589dcSJérôme Glisse * @size: size in bytes of the device memory to add 1401085ea250SRalph Campbell * Return: pointer to new hmm_devmem struct ERR_PTR otherwise 14024ef589dcSJérôme Glisse * 14034ef589dcSJérôme Glisse * This function first finds an empty range of physical address big enough to 14044ef589dcSJérôme Glisse * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which 14054ef589dcSJérôme Glisse * in turn allocates struct pages. It does not do anything beyond that; all 14064ef589dcSJérôme Glisse * events affecting the memory will go through the various callbacks provided 14074ef589dcSJérôme Glisse * by hmm_devmem_ops struct. 14084ef589dcSJérôme Glisse * 14094ef589dcSJérôme Glisse * Device driver should call this function during device initialization and 14104ef589dcSJérôme Glisse * is then responsible of memory management. HMM only provides helpers. 14114ef589dcSJérôme Glisse */ 14124ef589dcSJérôme Glisse struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 14134ef589dcSJérôme Glisse struct device *device, 14144ef589dcSJérôme Glisse unsigned long size) 14154ef589dcSJérôme Glisse { 14164ef589dcSJérôme Glisse struct hmm_devmem *devmem; 14174ef589dcSJérôme Glisse resource_size_t addr; 1418bbecd94eSDan Williams void *result; 14194ef589dcSJérôme Glisse int ret; 14204ef589dcSJérôme Glisse 1421e7638488SDan Williams dev_pagemap_get_ops(); 14224ef589dcSJérôme Glisse 142358ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 14244ef589dcSJérôme Glisse if (!devmem) 14254ef589dcSJérôme Glisse return ERR_PTR(-ENOMEM); 14264ef589dcSJérôme Glisse 14274ef589dcSJérôme Glisse init_completion(&devmem->completion); 14284ef589dcSJérôme Glisse devmem->pfn_first = -1UL; 14294ef589dcSJérôme Glisse devmem->pfn_last = -1UL; 14304ef589dcSJérôme Glisse devmem->resource = NULL; 14314ef589dcSJérôme Glisse devmem->device = device; 14324ef589dcSJérôme Glisse devmem->ops = ops; 14334ef589dcSJérôme Glisse 14344ef589dcSJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 14354ef589dcSJérôme Glisse 0, GFP_KERNEL); 14364ef589dcSJérôme Glisse if (ret) 143758ef15b7SDan Williams return ERR_PTR(ret); 14384ef589dcSJérôme Glisse 143958ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref); 14404ef589dcSJérôme Glisse if (ret) 144158ef15b7SDan Williams return ERR_PTR(ret); 14424ef589dcSJérôme Glisse 14434ef589dcSJérôme Glisse size = ALIGN(size, PA_SECTION_SIZE); 14444ef589dcSJérôme Glisse addr = min((unsigned long)iomem_resource.end, 14454ef589dcSJérôme Glisse (1UL << MAX_PHYSMEM_BITS) - 1); 14464ef589dcSJérôme Glisse addr = addr - size + 1UL; 14474ef589dcSJérôme Glisse 14484ef589dcSJérôme Glisse /* 14494ef589dcSJérôme Glisse * FIXME add a new helper to quickly walk resource tree and find free 14504ef589dcSJérôme Glisse * range 14514ef589dcSJérôme Glisse * 14524ef589dcSJérôme Glisse * FIXME what about ioport_resource resource ? 14534ef589dcSJérôme Glisse */ 14544ef589dcSJérôme Glisse for (; addr > size && addr >= iomem_resource.start; addr -= size) { 14554ef589dcSJérôme Glisse ret = region_intersects(addr, size, 0, IORES_DESC_NONE); 14564ef589dcSJérôme Glisse if (ret != REGION_DISJOINT) 14574ef589dcSJérôme Glisse continue; 14584ef589dcSJérôme Glisse 14594ef589dcSJérôme Glisse devmem->resource = devm_request_mem_region(device, addr, size, 14604ef589dcSJérôme Glisse dev_name(device)); 146158ef15b7SDan Williams if (!devmem->resource) 146258ef15b7SDan Williams return ERR_PTR(-ENOMEM); 14634ef589dcSJérôme Glisse break; 14644ef589dcSJérôme Glisse } 146558ef15b7SDan Williams if (!devmem->resource) 146658ef15b7SDan Williams return ERR_PTR(-ERANGE); 14674ef589dcSJérôme Glisse 14684ef589dcSJérôme Glisse devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; 14694ef589dcSJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 14704ef589dcSJérôme Glisse devmem->pfn_last = devmem->pfn_first + 14714ef589dcSJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1472063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 14734ef589dcSJérôme Glisse 1474bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; 1475bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1476bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1477bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1478bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1479bbecd94eSDan Williams devmem->pagemap.data = devmem; 1480bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 148158ef15b7SDan Williams 1482bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1483bbecd94eSDan Williams if (IS_ERR(result)) 1484bbecd94eSDan Williams return result; 14854ef589dcSJérôme Glisse return devmem; 14864ef589dcSJérôme Glisse } 148702917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add); 14884ef589dcSJérôme Glisse 1489d3df0a42SJérôme Glisse struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, 1490d3df0a42SJérôme Glisse struct device *device, 1491d3df0a42SJérôme Glisse struct resource *res) 1492d3df0a42SJérôme Glisse { 1493d3df0a42SJérôme Glisse struct hmm_devmem *devmem; 1494bbecd94eSDan Williams void *result; 1495d3df0a42SJérôme Glisse int ret; 1496d3df0a42SJérôme Glisse 1497d3df0a42SJérôme Glisse if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) 1498d3df0a42SJérôme Glisse return ERR_PTR(-EINVAL); 1499d3df0a42SJérôme Glisse 1500e7638488SDan Williams dev_pagemap_get_ops(); 1501d3df0a42SJérôme Glisse 150258ef15b7SDan Williams devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); 1503d3df0a42SJérôme Glisse if (!devmem) 1504d3df0a42SJérôme Glisse return ERR_PTR(-ENOMEM); 1505d3df0a42SJérôme Glisse 1506d3df0a42SJérôme Glisse init_completion(&devmem->completion); 1507d3df0a42SJérôme Glisse devmem->pfn_first = -1UL; 1508d3df0a42SJérôme Glisse devmem->pfn_last = -1UL; 1509d3df0a42SJérôme Glisse devmem->resource = res; 1510d3df0a42SJérôme Glisse devmem->device = device; 1511d3df0a42SJérôme Glisse devmem->ops = ops; 1512d3df0a42SJérôme Glisse 1513d3df0a42SJérôme Glisse ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, 1514d3df0a42SJérôme Glisse 0, GFP_KERNEL); 1515d3df0a42SJérôme Glisse if (ret) 151658ef15b7SDan Williams return ERR_PTR(ret); 1517d3df0a42SJérôme Glisse 151858ef15b7SDan Williams ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, 151958ef15b7SDan Williams &devmem->ref); 1520d3df0a42SJérôme Glisse if (ret) 152158ef15b7SDan Williams return ERR_PTR(ret); 1522d3df0a42SJérôme Glisse 1523d3df0a42SJérôme Glisse devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; 1524d3df0a42SJérôme Glisse devmem->pfn_last = devmem->pfn_first + 1525d3df0a42SJérôme Glisse (resource_size(devmem->resource) >> PAGE_SHIFT); 1526063a7d1dSDan Williams devmem->page_fault = hmm_devmem_fault; 1527d3df0a42SJérôme Glisse 1528bbecd94eSDan Williams devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; 1529bbecd94eSDan Williams devmem->pagemap.res = *devmem->resource; 1530bbecd94eSDan Williams devmem->pagemap.page_free = hmm_devmem_free; 1531bbecd94eSDan Williams devmem->pagemap.altmap_valid = false; 1532bbecd94eSDan Williams devmem->pagemap.ref = &devmem->ref; 1533bbecd94eSDan Williams devmem->pagemap.data = devmem; 1534bbecd94eSDan Williams devmem->pagemap.kill = hmm_devmem_ref_kill; 153558ef15b7SDan Williams 1536bbecd94eSDan Williams result = devm_memremap_pages(devmem->device, &devmem->pagemap); 1537bbecd94eSDan Williams if (IS_ERR(result)) 1538bbecd94eSDan Williams return result; 1539d3df0a42SJérôme Glisse return devmem; 1540d3df0a42SJérôme Glisse } 154102917e9fSDan Williams EXPORT_SYMBOL_GPL(hmm_devmem_add_resource); 1542d3df0a42SJérôme Glisse 15434ef589dcSJérôme Glisse /* 1544858b54daSJérôme Glisse * A device driver that wants to handle multiple devices memory through a 1545858b54daSJérôme Glisse * single fake device can use hmm_device to do so. This is purely a helper 1546858b54daSJérôme Glisse * and it is not needed to make use of any HMM functionality. 1547858b54daSJérôme Glisse */ 1548858b54daSJérôme Glisse #define HMM_DEVICE_MAX 256 1549858b54daSJérôme Glisse 1550858b54daSJérôme Glisse static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); 1551858b54daSJérôme Glisse static DEFINE_SPINLOCK(hmm_device_lock); 1552858b54daSJérôme Glisse static struct class *hmm_device_class; 1553858b54daSJérôme Glisse static dev_t hmm_device_devt; 1554858b54daSJérôme Glisse 1555858b54daSJérôme Glisse static void hmm_device_release(struct device *device) 1556858b54daSJérôme Glisse { 1557858b54daSJérôme Glisse struct hmm_device *hmm_device; 1558858b54daSJérôme Glisse 1559858b54daSJérôme Glisse hmm_device = container_of(device, struct hmm_device, device); 1560858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1561858b54daSJérôme Glisse clear_bit(hmm_device->minor, hmm_device_mask); 1562858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1563858b54daSJérôme Glisse 1564858b54daSJérôme Glisse kfree(hmm_device); 1565858b54daSJérôme Glisse } 1566858b54daSJérôme Glisse 1567858b54daSJérôme Glisse struct hmm_device *hmm_device_new(void *drvdata) 1568858b54daSJérôme Glisse { 1569858b54daSJérôme Glisse struct hmm_device *hmm_device; 1570858b54daSJérôme Glisse 1571858b54daSJérôme Glisse hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); 1572858b54daSJérôme Glisse if (!hmm_device) 1573858b54daSJérôme Glisse return ERR_PTR(-ENOMEM); 1574858b54daSJérôme Glisse 1575858b54daSJérôme Glisse spin_lock(&hmm_device_lock); 1576858b54daSJérôme Glisse hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); 1577858b54daSJérôme Glisse if (hmm_device->minor >= HMM_DEVICE_MAX) { 1578858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1579858b54daSJérôme Glisse kfree(hmm_device); 1580858b54daSJérôme Glisse return ERR_PTR(-EBUSY); 1581858b54daSJérôme Glisse } 1582858b54daSJérôme Glisse set_bit(hmm_device->minor, hmm_device_mask); 1583858b54daSJérôme Glisse spin_unlock(&hmm_device_lock); 1584858b54daSJérôme Glisse 1585858b54daSJérôme Glisse dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); 1586858b54daSJérôme Glisse hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), 1587858b54daSJérôme Glisse hmm_device->minor); 1588858b54daSJérôme Glisse hmm_device->device.release = hmm_device_release; 1589858b54daSJérôme Glisse dev_set_drvdata(&hmm_device->device, drvdata); 1590858b54daSJérôme Glisse hmm_device->device.class = hmm_device_class; 1591858b54daSJérôme Glisse device_initialize(&hmm_device->device); 1592858b54daSJérôme Glisse 1593858b54daSJérôme Glisse return hmm_device; 1594858b54daSJérôme Glisse } 1595858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_new); 1596858b54daSJérôme Glisse 1597858b54daSJérôme Glisse void hmm_device_put(struct hmm_device *hmm_device) 1598858b54daSJérôme Glisse { 1599858b54daSJérôme Glisse put_device(&hmm_device->device); 1600858b54daSJérôme Glisse } 1601858b54daSJérôme Glisse EXPORT_SYMBOL(hmm_device_put); 1602858b54daSJérôme Glisse 1603858b54daSJérôme Glisse static int __init hmm_init(void) 1604858b54daSJérôme Glisse { 1605858b54daSJérôme Glisse int ret; 1606858b54daSJérôme Glisse 1607858b54daSJérôme Glisse ret = alloc_chrdev_region(&hmm_device_devt, 0, 1608858b54daSJérôme Glisse HMM_DEVICE_MAX, 1609858b54daSJérôme Glisse "hmm_device"); 1610858b54daSJérôme Glisse if (ret) 1611858b54daSJérôme Glisse return ret; 1612858b54daSJérôme Glisse 1613858b54daSJérôme Glisse hmm_device_class = class_create(THIS_MODULE, "hmm_device"); 1614858b54daSJérôme Glisse if (IS_ERR(hmm_device_class)) { 1615858b54daSJérôme Glisse unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); 1616858b54daSJérôme Glisse return PTR_ERR(hmm_device_class); 1617858b54daSJérôme Glisse } 1618858b54daSJérôme Glisse return 0; 1619858b54daSJérôme Glisse } 1620858b54daSJérôme Glisse 1621858b54daSJérôme Glisse device_initcall(hmm_init); 1622df6ad698SJérôme Glisse #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ 1623